Support MuJoCo 3.12.0, widen pin to >=3.6,<3.13, and repair the skipped/stale test suites

[M-Colley]

Summary

Adds support for MuJoCo 3.12.0 and widens the pin to mujoco>=3.6,<3.13.

The MuJoCo side turned out to be small — no API MyoSuite uses was removed or changed incompatibly in 3.12. Most of this PR is fixing pre-existing NumPy 2 breakage and three test suites that were silently not testing what they claimed to.

MuJoCo 3.12

pyproject.toml pin bumped in both the core dependencies and the mjx extra, and uv.lock regenerated (mujoco 3.6.0 → 3.12.0, mujoco-mjx 3.6.0 → 3.12.0, warp-lang 1.11.1 → 1.16.0).

Nothing in the 3.12 breaking-change list applies: there is no flex, dcmotor, custom binary texture, mjx.render, or efm_L_* usage anywhere in the tree, and every mjtDyn/mjtGain reference is symbolic, so the enum renumbering is a no-op. Trajectories are effectively unchanged — max |Δqpos| ≈ 5e-7 over 50 steps with identical contact counts — so no policy retraining is implied.

The pin is widened rather than moved to 3.12-only: the full suite passes on 3.6.0 through 3.12.0, so there is no reason to force downstream users to upgrade.

NumPy 2 fixes (pre-existing, independent of MuJoCo version)

Named accessors like geomadr, vertadr, qposadr and dofadr return shape-(1,) arrays, and NumPy ≥2.3 no longer converts those to scalars implicitly. Nine sites:

• bimanual_v0 — geomadr/vertadr/vertnum/qposadr/dofadr indexed before use
• soccer_v0, chasetag_v0 — float(obs_dict["time"]) via .item()
• soccer_v0 — dropped size=(1,) from the joint-noise draws so a scalar is assigned into qpos[i] (RNG stream consumption is unchanged, so seeded reproducibility is preserved)

These reproduce identically on MuJoCo 3.6.0, so they are not 3.12 regressions. They only surface on Python 3.11+, where NumPy resolves to 2.4+; CI pinned 3.10, which resolves NumPy 2.2.6 and still permits the conversion as a deprecation — which is why they were invisible.

Also declares scipy, imported by bimanual_v0 and tabletennis_v0 but previously reaching CI only as an incidental scikit-video dependency.

Test suites

test_mjx.py — import mjx can never succeed (mjx ships as mujoco.mjx), so MJX_AVAILABLE was always False and all four tests silently skipped. Importing it correctly exposed API rot: mjx.device_put was removed (→ mjx.put_model), mjx.step takes (model, data) with control carried on Data, and the "xpos should not be all zeros" assertion is false for myofinger_v0, whose four bodies all sit at the origin.

test_editor.py (62 failures → 12 passed / 70 subtests) — the phalanx tests looked up thumbprox, 2proxph, … as body names, but those are geom/mesh names; the bodies are proximal_thumb, proxph2, … edit_fn_arm_reaching rebuilds each phalanx under its original body name, so every lookup returned None. test_digit_bodies_are_removed could never hold either, since the bodies are removed and re-added under the same name — replaced with test_digit_bodies_are_simplified, asserting what the edit actually does.

test_heightfields.py — constructed ChaseTagField/TrackField with a stale sim= wrapper, but HeightField takes mj_model/mj_data; ChaseTagField was also given one of its three required ranges.

CI

Adds a Python matrix (3.10, 3.12, 3.14) across the three runners with fail-fast: false. Keeping 3.10 matters — it is the floor declared by requires-python and the only tested interpreter resolving NumPy 2.2.6, the case that hid the breakage above. requires-python raised to <3.15 with a 3.14 classifier; MuJoCo 3.12 ships cp314 wheels.

mink

mink[examples] referenced an extra dropped in mink 1.0.0, which made uv lock warn. The requirement is now split by interpreter, because mink 1.3.0 — the first release without a <3.14 cap — also raises its floor to mujoco>=3.10.0, and requiring it unconditionally would drag every .[examples] install up to mujoco 3.10 and quietly invalidate the >=3.6 floor. Below 3.14 mink is left unconstrained; on 3.14 mujoco 3.6 has no wheels anyway (cp310–cp313 only).

Verification

test_myo (392 registered envs), test_editor and test_heightfields, all green:

MuJoCo (py3.13)	3.6.0	3.7.0	3.8.1	3.9.0	3.10.0	3.11.0	3.12.0
	OK	OK	OK	OK	OK	OK	OK

Python (mujoco 3.12.0)	3.10	3.11	3.12	3.13	3.14
numpy	2.2.6	2.4.6	2.5.2	2.5.2	2.5.2
	OK	OK	OK	OK	OK

MJX: 57 passed (the 4 previously-skipped test_mjx tests plus tests/mjx/).

Also verified that py3.10 + .[examples] + mujoco==3.6.0 still resolves and passes, so the widened floor is real rather than nominal.

Known behavioural change

3.12 gives built-in primitives canonical UV parameterisations (cc7fb98c). In this repo the visible effect is confined to the myoarm_bionic_bimanual pillars — cylinders carrying builtin="checker" mark="random" textures, which change from smooth vertical striping to mottled noise (~19% of pixels in a representative view). Textured planes and meshes are unaffected (myoelbow floor: 0.01% of pixels, max delta 1). Physics is unchanged; this only matters for screenshots and any vision-based bimanual observations.

🤖 Generated with Claude Code

[cla-assistant]

All committers have signed the CLA.

[Vittorio-Caggiano]

@M-Colley thanks for your PR and interest in myosuite. We are going to make a new release which should support the latest mujoco too. Do you need this version of myosuite compatible with the latest mujoco? in case, can you join slack and we can discuss more about it 1:1?